Download Report

Toward User-specific Tracking by Detection of Human Shapes in Multi-Cameras
Chun-Hao Huang1 , Edmond Boyer2 , Bibiana do Canto Angonese1 , Nassir Navab1 , Slobodan Ilic1,3
1
Technische Universit¨at M¨unchen, 2 LJK-INRIA Grenoble Rhˆone-Alpes, 3 Siemens AG
{huangc,slobodan.ilic,bibiana.canto,navab}@in.tum.de, edmond.boyer@inria.fr
Abstract
Human shape tracking consists in fitting a template
model to temporal sequences of visual observations. It usually comprises an association step, that finds correspondences between the model and the input data, and a deformation step, that fits the model to the observations given
correspondences. Most current approaches find their common ground with the Iterative-Closest-Point (ICP) algorithm, which facilitates the association step with local distance considerations. It fails when large deformations occur, and errors in the association tend to propagate over
time. In this paper, we propose a discriminative alternative for the association, that leverages random forests to
infer correspondences in one shot. It allows for large deformations and prevents tracking errors from accumulating. The approach is successfully integrated to a surface
tracking framework that recovers human shapes and poses
jointly. When combined with ICP, this discriminative association proves to yield better accuracy in registration, more
stability when tracking over time, and faster convergence.
Evaluations on existing datasets demonstrate the benefits
with respect to the state-of-the-art.
1. Introduction
Visual shape tracking is the process of recovering temporal evolutions of a template shape using visual information,
such as image silhouettes or 3D points. It finds applications
in several domains including computer vision, graphics and
medical imaging. In particular, it has recently demonstrated
a good success in marker-less human motion capture (mocap). Numerous approaches assume a user-specific reference surface, and the objective is to recover the skeletal
poses [29], surface shapes [6], or both simultaneously [15].
Most of these model-based methods [6, 12, 15, 18, 29]
can be viewed as extensions of Iterative-Closest-Point (ICP)
framework [5], which attempts to explain newly observed
data using the previous outcomes. As long as the initialization is close to the optimum solution, it is able to produce
(a) ref. surface
(b) input data
(c) detected shape
(d) results
Figure 1. Given a reference surface (a), our method discovers reliable data-model correspondences by random forests, color-coded
in (c). Instead of tracking, this strategy detects user-specific shapes
in a frame-wise manner, resulting in more sustainability.
outstanding results. However, they also suffer from inherent weaknesses of generative strategies, e.g., the slow convergence. Moreover, when large deformations or outliers
occur, discovering associations between data and models is
particularly difficult. Unreliable associations result in ambiguous situations that yield erroneous numerical solutions
and, consequently, break the tracking process.
In contrast, discriminative approaches that ‘detect’ rather
than track models have shown better robustness over
the past decade, for instance, in human pose estimation
with Kinect [22, 27]. These approaches operate frameindependently, and are generally drift free. In this paper,
we explore this direction in order to get robust observationmodel associations, regardless of the results from previous
frames. We further present a discriminative ‘tracking-bydetection’ human mocap framework, as in Fig. 1. Inspired
by Taylor et al. [27], we apply regression forests to improve the associations. Shape geometries are characterized
by volumetric representations, and are fed into user-specific
forests to predict correspondences in one shot. Contrary to
generative methods, this prediction does not require close
initializations from a nearby frame. In addition, it allows
a single model to be used as a reference surface for several
different sequences, again even if large deformations or outliers exist. We combine this strategy with a generative tracking approach that takes our one-shot associations as input.
Experiments demonstrate that this hybrid discriminativegenerative framework leads to better, or comparable results
Volumetric
normal field
Regression
forest
NN search
𝑡
deformed ℳ
voxelization
Joint Shape and
Pose Optimization
Ref. surface ℳ
𝒴
predicted corres.
𝒴𝑡
input data
𝑡
Figure 2. Pipeline of our framework. Correspondences are visualized in the same color. Black means no correspondence for that data point.
than purely generative approaches, e.g. [6,8], reducing error
accumulations and hence increasing the stability.
The contribution of this work is two-fold: (1) a one-shot
correspondence inference with complete 3D input (rather
than 2.5D as in [22,27]), which leads to (2) a hybrid subjectspecific human pose and shape capture method that relies
little on former results (unlike other ICP-like methods [6,
12, 15, 18]), holding the ability to recover from drifting.
2. Related work
Among the vast literature on human pose estimation
[19], we focus on top-down approaches that assume a 3D
model and deform it according to input data, either directly
with pixels as in [12,18,25], or with 3D points as in [6,8,15].
These methods typically decompose into two main steps:
(i) data association, where observations are associated to
the model, and (ii) deformation estimation, where deformation parameters are estimated given the associations. Our
primary objective in this paper is to improve the first part.
Existing approaches for this issue are discussed below.
Generative approaches. Methods in this category do not
require any training. They follow the association strategy
in ICP while extending the motion model to more general
deformations than the one in the original method [5]. Associations are addressed by searching for closest points, with
various distance measures such as point-to-point [6], pointto-plane [7], or Mahalanobis distances [25]. This strategy
heavily relies on the fact that observations in consecutive
frames are in vicinity. Klaudiny et al. [17] generalize the
idea from the previous frame to a certain frame in the considered sequences, finding the best non-sequential order to
track, but the proximity assumption remains. Gall et al. [12]
establish 3D-2D correspondences by considering both texture in images, and contours in silhouettes. Later, Liu et
al. [18] include image segmentation information in order
to differentiate multiple interacting subjects. These approaches implicitly assume that observations only describe
the tracking subjects, which does not necessarily hold in 3D
data that often contain fake geometries. Cagniart et al. [6]
introduce an additional outlier class to reject associations
with noisy observations. Data is explained by Gaussian
Mixture Models (GMM) in an Expectation-Maximization
(EM) manner. Huang et al. [15] follow a similar concept,
but aggregate the outlier likelihood over every Gaussian
component and offer better robustness. All these generative methods are highly likely to fail in large deformations.
Furthermore, they are prone to error accumulations and,
as a result of matching several successive frames wrongly
(whether sequentially or not), they are prone to drift.
Discriminative approaches. Recently, discriminative approaches have demonstrated their strengths in tracking human poses with depth images [4, 27]. Taylor et al. [27] propose a single-frame, or so called one-shot strategy, which
yields decent dense correspondences without iterative refinements. With the help of regression forests, they map
each foreground pixel to a weighted point in 3D, and
thereby search the closest point within a predefined surface. Later, Pons-Moll et al. [20] train forests with a new
objective in metric space, and couple the one-shot strategy with ICP. In the case of full 3D, Kanaujia et al. [16]
use shape context histograms as descriptors, segment visual hulls into body parts with a pre-trained Support Vector
Machine (SVM), and build the skeletons with the method
similar to [22]. Starka et al. [23] formulate the matching
problem as the inference of Markov random field (MRF).
Rodola et al. [21] apply forests to learn the parameters of
wave kernel signatures [3] during training, and facilitate
dense matching between two meshes.
To avoid computation overhead, we develop volumetric
normal fields to describe meshes in a discretized volume
Ω3 , and extend the comparison features from 2.5D [22, 27]
to full 3D data. Our method can be viewed as a trackingby-detection approach for human shape tracking.
3. Preliminaries and method overview
We first state the problem and briefly outline our method.
A 3D reference surface is denoted as M = (M, TM ),
3
v
where M = {xv }N
v=1 ⊂ R are the locations of vertices
v, and TM defines the triangles. Evolving M typically
amounts to parameterizing M as a function of deformation
parameters Θ, namely, M(Θ). Unlike [12, 18, 27], our interest lies in recovering not only poses but also shapes of
the subjects. We adopt a surface deformation framework
that groups vertices into Np patches [6], and assign each of
(3)
𝐞𝟐
(4) 𝐞𝟐
(a)
them a rigid body motion θ. Thus, Θ is the collection of all
Np
, encoding the global shape of the surface.
θ, Θ = {θk }k=1
We refer the readers to [6] for detailed explanations.
Given an observed visual hull Y t = (Yt , TYt )1 , where
Ny
Yt = {yi }i=1
⊂ R3 , the goal is to determine the optit
ˆ such that Mt = M(Θ
ˆ t ) resembles Yt as much as
mal Θ
possible. It typically boils down to two sub-problems:
1. finding correspondence pairs C = {(i, v)} between the
vertex sets of Y and the vertex sets of M2 , and
2. minimizing an energy E that describes the discrepanˆ = arg minΘ E(Θ; C).
cies between vertices in C: Θ
Standard ICP-based approaches [6, 8] alternate between
these two steps, refining C t and Θt iteratively. They require
Mt−1 to be close to Yt , and is usually slow to converge.
We develop a different strategy that warps the input data
˜ TY ), and
Y to the reference mesh M, denoted as Y˜ = (Y,
visualized as a triangular mesh in Fig. 2. If the warping is
perfect, this mesh will look clean and resemble M as much
as possible. Incorrect mapping, on the other hand, can be
˜ represent the lotold from huge edges. Vertex positions Y
cations of potential matches between Y and M. Thus, C can
be built directly by doing nearest neighbor search between
˜ and M just once, as illustrated in Fig. 2.
Y
Specifically, we consider this R3 → R3 mapping as
a composite one: R3 → Ω3 → R3 . The former mapping is voxelization (Sec. 4.1), while the latter is regression
(Sec. 5). A forest is trained with many voxelized meshes
off-line (Sec. 5.1). During runtime, yi is first mapped to a
˜ Given
˜ i ∈ Y.
voxel vi , and then regressed to a 3D point y
the properly estimated C, poses and shapes are then recovered as in [14] (Sec. 5.2). Fig. 2 shows our pipeline.
4. Normal volume and features
Before entering the forest, we cast our data into a volumetric field N : Ω3 ⊂ R3 → R3 , and design a set of
features to describe surface geometries in volumes.
4.1. Volumetric normal field (VNF)
Voxelizing a mesh in general consists of two parts: (1)
finding voxel positions for every vertex, and (2) testing the
overlap between triangles and voxels. The first part can be
viewed as a quantization mapping from Euclidean space to
a discretized space v : R3 → Ω3 . The size of the volume is large enough to include all possible pose variations,
and its center is aligned with the mean of the surfaces. The
voxel size is chosen to be close to the average edge length
1 Both the reference model and input data are described as triangular
meshes. Although the terms mesh and surface are used interchangeably in
the text, we refer only to input data as visual hulls.
2 Whenever it is clear from the context, we will drop the time dependent
variable t, in order to keep notations uncluttered.
(a)
(b)
(b)
(c)
Figure 3. (a) original offset pair ψ. (b) η = 0 results in ψ without
re-orientation, i.e. R = I. (c) η = 1. ψ is re-orientated by a
rotation matrix R = [e1 , e2 , e3 ] characterized by a LCF.
of meshes, so that a single voxel is not mapped by too many
vertices. To check the intersection of triangles with voxels, we apply separating axis theorem which is known to be
efficient for collision detection in graphic community [1].
After determining voxels occupied by the surface, referred to as vsuf , we further identify voxels located inside
and outside the surface, denoted respectively as vin and vout .
Together they define a truncated normal field:


if vout
+
3
(1)
N(v) = n ∈ [−1, +1] , knk = 1 if vsuf


−
if vin .
n is the average surface normal from the containing triangles, whereas is an indicator vector ( [2, 2, 2] in our
experiments). Given a mesh, our VNF naturally encodes
both spatial occupancies (surface locations) and first-order
derivatives (normals). It shares a similar spirit with implicit
surface representations, e.g. level-set, but does not require
any expensive distance transform computations.
4.2. Volumetric features
Here, we present the features f (vsuf ) for describing the
above VNF, which are later used to train the forests. Conventionally, only one feature dimension κ is selected to
separate data at each branch node of the forest [9]. One
does not have to prepare the whole high-dimensional feature vector for predictions, because only a few dimensions
are needed. The calculation of f is suggested to be dimensionally independent. We therefore avoid descriptors that
requires normalization, like MeshHOG [30], or SHOT [28],
and resort to comparison features used in [10, 22].
As depictcted in Fig. 3, for each surface voxel vsuf (blue),
we shoot two offsets (red vectors) ψ = (o1 , o2 ) ∈ Ω3 ×
Ω3 , and reach two neighboring voxels (green). To describe
the local geometry, we concatenate the dot product of their
normals, and the difference of VNF within local cuboids to
form the feature vector f . Such a computation is repeated
for multiple offsets at vsuf . By definition, each dimension
of f can be evaluated independently. See Supplementary
Material for more details.
Note that f is a function of vsuf , but takes an offset pair ψ,
a binary variable η (whether using Local Coordinate Frame
𝐧
5.1. Random forest
𝝋
𝒅
(1)
(2)
𝐧
A forest is an ensemble of T binary decision trees, each
separating data with split functions at branch nodes and
storing statistic models at leaf nodes. The training objectives, split functions, and statistical models vary from task
to task. We refer readers to [9] for a comprehensive review.
𝐞𝟏 = 𝐧
𝐞𝟑
𝐞𝟐
(3)
(4)
(a)
𝐞𝟐
(b)
Figure 4. (a) The way we compute LCF. We search neighboring
surface voxels to get a reference vector ( purple arrow); see text for
details. (b) Our method leads to LCFs covariant to pose variations.
(LCF) or not), and a rotational matrix R ∈ SO(3) (the orientation of LCF) as parameters, denoted as f (vsuf ; Rη (ψ)).
η determines the alignment of the offset ψ w.r.t a LCF,
whose transformation is specified by R. The intuition behind this adjustment is to make features f invariant to poses,
(a) 3(b-c). Without(b)
(c) land on
c.f. Fig.
re-orientations, ψ might
different types of voxel pairs (c.f. Fig. 3(a-b)), and hence
cause different feature responses, despite the fact that the
current voxels are located on the same position on the body.
Both offset pairs ψ and binary variables η are learned when
training the forest, while the rotational matrix R is characterized by a local coordinate frame obtained as follows.
Local coordinate frame. Constructing a LCF requires 3
orthonormal vectors: [e1 , e2 , e3 ]. They usually come from
doing PCA on a local region, which is impractical to repeat
for all surface voxels vsuf . Instead, we apply the following
method, as illustrated in Fig. 4(a). The first eigenvector e1
is always the normal vector (red arrow). e3 is simply e1 ×e2
if the second eigenvector e2 is known. First we open a local
cuboid around the current voxel (blue). Among neighboring gray surface voxels, we find the one that has least normal changes, and yet located far from the current voxel. In
practice, given the notations in Fig. 4(a), we select the one
that maximizes (dˆ + cos ϕ), where ϕ is the angle between
normals, and dˆ is the normalized distance. Given this reference voxel (yellow), a reference vector (purple) is decomposed into two components. The component perpendicular
to the normal is then chosen as e2 (green). Fig. 4(b) shows
that this approach produces LCFs quasi-covariant to pose
changes, and as a result, quasi-pose-invariant features f .
5. Correspondence inference and tracking
Now we proceed to the second mapping: Ω3 → R3 ,
where the surface voxel vsuf traverses a pre-trained regression forest according to the feature responses, and attains a
point in R3 that lies on the surface embedding defined by
the vertices of the reference model M.
Training. To incorporate abundant training variations, we
animate M to a variety of poses with a method similar
to [26]. After voxelizing all the animated meshes, we associate each surface voxel to their locations at rest pose, and
obtain a pool of sample-label pairs S = {(vsuf , xv )}. Each
tree is trained with a random subset of S. Our splitting candidate is φ = (ψ, η, κ, τ ) (offset pairs ψ, binary variables
η, testing feature channels κ, and thresholds τ ). Let SN
denotes the samples arriving at a certain branch node. The
training process is to partition SN recursively into two subsets SL and SR , based on randomly generated φ:
SL (φ) = {vsuf ∈ SN |fκ (vsuf ; Rη (ψ)) ≥ τ }
(2a)
SR (φ) = {vsuf ∈ SN |fκ (vsuf ; Rη (ψ)) < τ }.
(2b)
Here fκ denotes the κth dimension of the feature f .
Whether φ is a good split or not depends on whether it
produces more homogeneous subsets. In supervised learning, this is often measured by information gain:
G(φ) = H(SN ) −
X
i∈{L,R}
|Si (φ)|
H(Si (φ)),
|SN |
(3)
where H = σ 2 (·) is the entropy, measured by the variance
of all xv in the sample set as in [13]. The split that maximizes the information gain, φ∗ = arg maxφ G(φ), is stored
for the later prediction use. The tree recursively splits samples and grows until one of the following stopping criteria is
true: (1) it reaches the maximum depth, or (2) the number of
samples |SN | is too small. Once a leaf node is reached, we
perform mean-shift clustering to represent the distributions
of xv as a set of confidence-weighted modes H = {(h, ω)}.
h ∈ R3 is the mode location and ω is a scalar weight.
Outliers such as false geometries, or un-removed background often exist in visual hulls, drastically deteriorating
tracking results. If their models are available, we also include them in the training process, so that forests can reject
them online. In this case, the goodness of a split φ should
be evaluated in terms of both classification and regression.
We follow Fanelli et al. [11] and adapt the entropy as:
X
D
H(S) = −
p(c|S) log p(c|S) + (1 − e α )σ 2 (S), (4)
c
where p(c|S) is the class probability of being foreground
or background. Eq. 4 is the weighted sum of the afore-
mentioned regression measure σ 2 and the classification entropy measure. The regression part gets increasing emphasis when the current depth D gets larger (i.e., the tree grows
deeper), and the steepness is controlled by the parameter α.
𝑖
…
𝒩𝑖
𝒢𝑖
𝒢𝒩𝑖
Subject coordinate frame. To achieve global rotation invariance, we always rotate meshes into a canonical orientation before the voxelization. In the training phase, since
each animated mesh is accompanied with a skeletal pose,
we compute two unit-length vectors from the skeleton, and
align them to x- and z-axis respectively. The first one is
the common perpendicular vector of bone RShoulder-Torso
and bone LShoulder-Torso, while the second one is the sum
of them. See Supplementary Material for illustrations.
Recall that the volume center is aligned with the means
of meshes, which brings translation invariance. Together
these two steps characterize a subject coordinate system,
simulating the PCA process on a whole mesh. During tracking, we approximate the skeletal pose using results of the
previous frame, and repeat the same process.
Prediction. In the prediction phase, a voxel lands on T
leaves containing different collections of modes, denoted
together as G = {H1 ...HT }. A standard way of aggregation is doing mean-shift and keeping the cluster with largest
weight. This is usually done independently for each sample. For instance, in Fig. 5(a), the green vertex aggregates
with only green leaves, and black vertices aggregate with
their respective black leaves as well. The consequent Y˜ is
however, often noisy, where vertices of same triangles are
mapped to locations distant from each other, as in Fig. 5(b).
We utilize the triangles TY to ameliorate this problem.
All yi are first mapped to voxels, sent into the forest once,
and return in total Ny × T predictions. For each vertex
i, we consider not only its own predictions, but also those
from the neighbors Ni , e.g. all green and black leaves in
Fig. 5(a). We sort the set Gi ∪GNi descendingly according to
their weights ω, and do mean-shift only with the first half of
them. The new mode location with highest confidence is the
˜ i . This strategy respects the mesh connectivity
final output y
and results in more structured forest predictions. Compare
Fig. 5(c) to (b) to see the improvements.
Nearest neighbor search. Given the regression results
˜ i , each vertex i in input gets a closest vertex p in the refy
erence vertex set VM , i.e. p = arg minv∈VM k˜
yi − xv k2 .
Similar to [6], we reject the searched correspondences if
their normals differ from each other too much. Our advantage over [6] is that M and its normals are fixed throughout
tracking. One does not have to re-compute normals online,
and other speed up algorithms like kd-tree are also feasible.
Each correspondence pair (i, p) is associated to a weight
wip = exp(−d2 (i, p)/2l2 ), where d(·) stands for Euclidean
…
(a)
(b)
(c)
Figure 5. (a) Traditional approach aggregates leaves of each vertex (green) independently, while we also take the 1-ring neighbors
(black) into account. (b) Forest output Y˜ by aggregating each Gi
separately. (c) Forest output Y˜ by aggregating Gi ∪ GNi .
distance, and l is the averaged edge length on M. Since the
forest gives relatively good initial guess of the matches. It
is much easier to find the right matches in this way.
5.2. Energy functions
Recall that our goal is to estimate deformation parameter
Θ, whose resulting M(Θ) best explains Y. Given correspondences C from above, a data term is formulated as:
X
Edata (Θ; C) =
wip kyi − xp (Θ)k22 ,
(5)
(i,p)∈C
which is a standard sum of weighted squared distances.
Since evolving a surface with discrete observations (even
with a good C) is ambiguous by nature, regularization terms
are usually introduced to exert soft constraints. We follow
the simultaneous estimation framework [14], which applies
the following regularization terms:
Preserving local rigidity. Without any prior knowledge
of motion, patches are preferred to be distributed uniformly
on M. Given a vertex v, the rigidity constraint enforces the
predicted positions xv (θk ) and xv (θl ) from two adjacent
patches Pk and Pl ∈ Nk to be consistent:
Er (Θ) =
NP X
X
X
wkl kxv (θk ) − xv (θl )k22 , (6)
k=1 Pl ∈Nk v∈Pk ∪Pl
where Θ is implicitly encoded in xv (θk ) and xv (θl ).
Pose posteriors. It is noteworthy that the method described so far applies to not only humans but all types of
surfaces. Nevertheless, since our application is capturing
human motions, skeletal poses are also of our interest. We
employ the skeleton binding energy that keeps the relationship between skeletons and surfaces:
Eb (Θ, J ) =
NP
X
wk kTθk (β 0k ) − β k k22 .
(7)
k=1
Here β ∈ R3 stands for the β-coordinate proposed by
Straka et al. [25]. Θ is encoded in Tθk ∈ SE(3). Our
Sequence
Views
Frames
48
8
8
8
8
9
9
9
9
176
173
149
173
174
91
130
93
148
Goalkeeper [2]
Crane [29]
Jumping [29]
Handstand [29]
Bouncing [29]
Cutting
WalkChair1
HammerTable [15]
WalkChair2 [15]
Outliers
Err. metric
Compared approaches
Subject / # Vertices
√
√
√
√
A
A
A
A
A
A
A&B
A&B
-
S1 / 4980
S2 / 3407
surICP [6], artICP [8]
S3 / 3848
fixOL [6], bpSVM [14],
patchedOL [15]
[6], [14], [15],
[24] + [25]
S4 / 5211
S5 / 5233
Table 1. Sequences used for evaluation. We apply two different error measures, depending on the provided ground truths. A: silhouette
overlap errors averaged over all views. B: distances to annotated joint positions in pixels.
% of matches within
max allowed distance
100%
80%
75.47%
60%
40%
48.39%
VNF-forest
MeshHOG-1NN
20%
0%
0.000
0.050
0.100
0.150
0.200
0.250
0.300
Max allowed geodesic distance to ground truth
Figure 6. Matching error comparison of our method and MeshHOG of S2. We yields more locally consistent correspondences.
N
j
pose parameter J = {xj }j=1
is a set of 3D locations for
Nj = 15 joints, implicitly encoded in β.
Combining Eq. 5-7, we have the final energy:
E(Θ, J ; C) = λEdata (Θ; C) + Er (Θ) + Eb (Θ, J ). (8)
λ defines the softness of the surface, and is empirically set
as 10 throughout our experiments. Given an input Y, the re˜ and hence a fixed
gression forest returns a fixed response Y,
C. We therefore apply standard Gauss-Newton method directly to find the minimizer of Eq. 8. Note anyway that refining C like ICP is always possible. In this case, our method
provides better initializations than using last frame results,
reducing the number of needed ICP-iterations.
6. Experiment results
The proposed method was evaluated extensively on 9 sequences, whose profiles are summarized in Table 1. An individual forest is trained for each subject with up to 200
meshes, depending on the number of vertices per mesh. For
S1 − 3 we train standard regression forest; for S4 & S5 we
apply the adaptation in Eq. 4 due to the un-properly segmented chairs and tables in input data (α = 2). Growing
T = 20 trees to depth 25 with 15000 testing offset pairs
ψ takes about 8 hours. The performance of our method is
analyzed in different aspects, both qualitatively and quantitatively. We evaluate shapes with the widely-used silhouette
(a)
(b)
(c)
(d)
Figure 7. (a) references surface of S2 (vertex indices encoded with
colors). (b) test mesh. (c) matching results with MeshHOG. (d)
matching results of the proposed VNF-forest framework.
overlap error, and evaluate poses with the 2D distances between ground truths and the projected joint locations.
Descriptiveness. First we briefly demonstrate the discriminability of the VNF-forest combination. The task is
finding correspondences on the reference surface of S2. We
consider animated meshes as testing data, where groundtruth vertex indices are available. The error metric is the
geodesic distance between predicted vertices and groundtruths, and we compare with MeshHOG [30] which is the
extension of image-based histogram of oriented gradient
(HOG) on surface manifolds.
Fig. 6 depicts the proportion of vertices whose errors
are less than a certain distance. The VNF-forest framework gives overall 27.34% exact matches, while searching
the vertex with closest MeshHOG response gives 46.07%.
However, when increasing the tolerance of correctness, our
method presents a significant improvement, and yet MeshHOG gains only a little. With our method, 75.47% of vertices obtain correspondences no farther than three times averaged edge length (green dashed line), and 48.39% for
MeshHOG. Moreover, almost half of matches from MeshHOG lie outside the range of roughly the length of the lower
arm (orange dashed line) from the ground truths. Such a
phenomenon is confirmed in Fig. 7. Taking Fig. 7(b) as input, MeshHOG returns results as in Fig. 7(c), whereas our
method gives visually smoother matches like Fig. 7(d). The
proposed VNF-forest framework certainly provides more
ours + ICP
[24] + [25]
bpSVM [14]
patchedOL [15]
sil. error
sil. error
sil. error
sil. error
WalkChair1
7432
fail
8931
Cutting
4048
fail
20385
sil. error
joint error
sil. error
joint error
sil. error
joint error
sil. error
joint error
HammerTable
4019
13.3 ± 6.9
17285
64.2 ± 53.9
fail
fail
3593
10.1 ± 3.0
WalkChair2
6144
15.8 ± 6.1
12219
20.6 ± 22.0
18063
24.6 ± 10.7
6803
15.9 ± 6.3
Table 2. Pixel overlap error of 4 sequences over all frames and all cameras. Image resolution: 1000 × 1000.
source
surface
(a)
fixOL [6]
sil. error
24976
fail
sil. error
fail
18482
target points
(b)
(c)
(d)
Figure 8. (a) the source surface and target points. (b) registration
results (green) of EM-ICP [6] after 30 iterations. (c) results of
VNF-forest framework (no iterations). (d) results of VNF-forest
framework plus 14 iterations of EM-ICP.
locally consistent associations, which, as shown below, are
sufficiently accurate to guide plausible deformations.
Registration. Next we show the capacity of the proposed
method on registration tasks. With the source surface and
distant target points as in Fig. 8(a), ICP-like approach only
produces results like Fig. 8(b), where many discrepancies
still remain after 30 iterations. Instead, our VNF-forest
framework brings the source mesh closer to the target points
in one shot as in Fig. 8(c). After 14 ICP-iterations it refines
the results as in Fig. 8(d). In the following experiments if
reference surfaces are not aligned with the first frames, we
also register them before the tracking starts.
Tracking without outliers. For sequences without outliers, we compare with surface-based ICP (surICP) [6] and
articulated ICP (artICP) [8], both of which explain data with
GMM using the Expectation-Maximization algorithm. We
run an additional ICP step to reduce the errors (ours + ICP)
for all testing sequences. The averaged silhouettes overlap
errors are shown in Fig. 10(a-b). In general, our method
performs much better than artICP, and yields comparable
results with surICP. However, our method requires less ICPiterations to converge (see Supplementary Material). It follows that, compared with using results of previous frames
as initializations, our method is capable of providing better
ones. Examples of the estimated correspondences, the deformed surfaces and the skeletons are shown in Fig. 12(a).
We further investigate what happens when there are
tracking errors in previous frames. Fig. 9(a) shows the results of frame 91 in sequence bouncing. Note that left/right
(a)
(b)
(c)
(d)
Figure 9. (a) results of frame 91 in bouncing. (b) visual hulls of
frame 92 and correspondences from the forest in colors. (c) results
of frame 92 from our approach. (d) results from surICP [6].
legs are crossed due to their close interaction with each
other. In the next frame, when they are separate in the visual
hull, the forest discovers correspondences more correctly, as
in Fig. 9(b) (c.f. the colors of M in Fig. 2), and lead us to
results in Fig. 9(c) without self-intersections. ICP strategy
discover wrong associations and gives results in Fig. 9(d)
under the same softness parameter λ. Errors propagate to
the next frame, and gradually deteriorate the tracking, unless luckily regularization terms implies the opposite.
Tracking with outliers. Four of our testing sequences,
Cutting, WalkChair1, HammerTable, and WalkChair2 contain tables or chairs in observations, which play the roles
as static outliers. We compare with other outlier rejection strategies such as, fixed outlier proportion (fixOL) [6],
removing outliers by body-part classifications with SVM
(bpSVM) [14], and modeling outlier likelihood dynamically by aggregating over all patches (patchedOL) [15].
As shown in Fig. 10(c-d), conventional outlier strategy
fixOL drifts quickly and soon fail to track (green curves).
ICP with robust outlier treatment, patchedOL, is able to
sustain noisy input to a certain extent. Once it starts drifting, the error only gets higher due to its ICP nature (yellow curves). When subjects and outliers are sperate components in visual hulls, we cast them into VNF, and feed them
separately into the joint classification-regression forest. If
they are connected to each other, forests inevitably associate
some outliers to vertices on the model, and cause undesirable deformations, as the spike in blue curves in Fig. 10(d).
Nonetheless, since we rely less on previous frames for data
associations, the results can always get recovered when they
are separated again. In average, we still yield low errors
16000
14000
12000
10000
8000
6000
4000
0
20
40
60
80
100
120
140
160
180
29000
ours + ICP
Cagniart et al. (surICP)
Corazz et al. (artICP )
24000
19000
14000
9000
4000
20
40
60
10000
5000
0
30
40
ours
10
100
120
140
160
180
50
60
70
patchedOL
Straka et al.
1
21
41
bpSVM
120
100
20000
15000
10000
5000
ours + ICP
0
20
40
fixOL
60
81
101
121
141
(a) WalkChair2
25000
0
61
Time (frame)
30000
patchedOL
80
80
60
40
ours
patchedOL
Straka et al.
bpSVM
20
0
100
1
21
41
61
81
Time (frame)
Time (frame)
Time (frame)
(c) Cutting
80
average joint error
Silhouette overlap error (pix.)
Silhouette overlap error (pix.)
15000
20
15
(b) Bouncing
ours + ICP
Cagniart et al. (fixOL)
Huang et al. (patchedOL)
10
20
Time (frame)
30000
0
25
0
0
(a) Crane
20000
30
5
Time (frame)
25000
35
average joint error
18000
Silhouette overlap error (pix.)
Silhouette overlap error (pix.)
ours + ICP
Cagniart et al. (surICP)
Corazz et al. (artICP )
20000
(b) HammerTable
(d) WalkChair1
Figure 10. Pixel overlap error of 4 sequences, averaged over all cameras. Image resolution: Figure 11. Averaged joint errors. Image resolution: 1000 × 1000.
(a-b): 1920 × 1080. (c-d): 1000 × 1000
(a)
(b)
(c)
throughout the whole sequences, as in Table 2. We remark
that such ability to recover is the essence of our discriminative approach, which is the biggest advantages over the
existing generative methods.
We also verify the efficacy on pose estimations, and
compare to another single-frame approach from Straka et
al. [24, 25]. The error metric is the discrepancy between
estimated joint positions and the annotated ground truths.
Fig. 11 plots the results of sequences HammerTable and
WalkChair2, where we confirm again considerably better accuracy than bpSVM [14] (red curves), Straka et
al. [24, 25] (green curves), and comparable results with
patchedOL [15]. The recovered shapes and poses are also
presented in Fig. 12(c), superimposed on original images.
7. Conclusion
deformed surface
Figure 12. Qualitative results. (a) input visual hulls and the estimated correspondences visualized in different colors. Black color means
no correspondence found for that vertex. Top: Handstand; bottom: jumping. (b) estimated shapes and poses of Goalkeeper. (c) estimated
(d) on images. Blue: our results. Yellow: Cagniart et al. [6].
shapes and poses of Cutting, WalkChair2, and WalkChair1, overlayed
We present a hybrid human shape tracking approach
equipped with an alternative data-model association strategy. With the help of regression forests, we learn from preanimated meshes to discover better correspondences on input visual hulls. The one-shot property effectively prevent
tracking errors from accumulating, yielding more stability
compared with other generative ICP extensions. It contributes to locally consistent correspondences, which speeds
up the convergence of ICP when used as the initializations.
The reliability of the proposed method is confirmed by the
experiments on numerous public sequences. Future directions include alleviating problems of topological changing
and incorporating photometric information.
Acknowledgment. This work is funded by Taiwan Ministry of Education Scholarship n°1001121US089, and partially funded by Technische Universit¨at M¨unchen - Department of Informatics - Chair for Computer Aided Medical
Procedures & Augmented Reality.
References
[1] T. Akenine-M¨oller. Fast 3d triangle-box overlap testing. In SIGGRAPH 2005 Courses. ACM, 2005.
[2] B. Allain, J.-S. Franco, E. Boyer, and T. Tung. On
Mean Pose and Variability of 3D Deformable Models.
In ECCV. Springer, 2014.
[3] M. Aubry, U. Schlickewei, and D. Cremers. The wave
kernel signature: A quantum mechanical approach to
shape analysis. In ICCV Workshops. IEEE, 2011.
[4] A. Baak, M. Muller, G. Bharaj, H.-P. Seidel, and
C. Theobalt. A data-driven approach for real-time
full body pose reconstruction from a depth camera. In
ICCV. IEEE, 2011.
[5] P. J. Besl and N. D. McKay. A method for registration
of 3-d shapes. PAMI, 1992.
[6] C. Cagniart, E. Boyer, and S. Ilic. Probabilistic deformable surface tracking from multiple videos. In
ECCV. Springer, 2010.
[7] Y. Chen and G. Medioni. Object modeling by registration of multiple range images. In International
Conference on Robotics and Automation. IEEE, 1991.
[8] S. Corazza, L. M¨undermann, E. Gambaretto, G. Ferrigno, and T. P. Andriacchi. Markerless motion capture through visual hull, articulated icp and subject
specific model generation. IJCV, 2010.
[9] A. Criminisi and J. Shotton. Decision forests for computer vision and medical image analysis. Springer,
2013.
[10] A. Criminisi, J. Shotton, D. Robertson, and
E. Konukoglu.
Regression forests for efficient
anatomy detection and localization in ct studies. In
Medical Computer Vision. Recognition Techniques
and Applications in Medical Imaging. Springer, 2011.
[11] G. Fanelli, M. Dantone, J. Gall, A. Fossati, and
L. Van Gool. Random forests for real time 3d face
analysis. IJCV, 2013.
[12] J. Gall, C. Stoll, E. De Aguiar, C. Theobalt, B. Rosenhahn, and H.-P. Seidel. Motion capture using joint
skeleton tracking and surface estimation. In CVPR.
IEEE, 2009.
[13] J. Gall, A. Yao, N. Razavi, L. Van Gool, and V. Lempitsky. Hough forests for object detection, tracking,
and action recognition. PAMI, 2011.
[14] C.-H. Huang, E. Boyer, and S. Ilic. Robust human
body shape and pose tracking. In 3DV, 2013.
[15] C. H. Huang, E. Boyer, N. Navab, and S. Ilic. Human
shape and pose tracking using keyframes. In CVPR.
IEEE, 2014.
[16] A. Kanaujia, N. Kittens, and N. Ramanathan. Part segmentation of visual hull for 3d human pose estimation.
In CVPR Workshop. IEEE, 2013.
[17] M. Klaudiny, C. Budd, and A. Hilton. Towards optimal non-rigid surface tracking. In ECCV, 2012.
[18] Y. Liu, C. Stoll, J. Gall, H. P. Seidel, and C. Theobalt.
Markerless motion capture of interacting characters
using multi-view image segmentation. In CVPR.
IEEE, 2011.
[19] T. B. Moeslund, A. Hilton, and V. Kr¨uger. A survey of
advances in vision-based human motion capture and
analysis. CVIU, 2006.
[20] G. Pons-Moll, J. Taylor, J. Shotton, A. Hertzmann,
and A. Fitzgibbon. Metric regression forests for human pose estimation. In BMVC, 2013.
[21] E. Rodola, S. R. Bulo, T. Windheuser, M. Vestner, and
D. Cremers. Dense non-rigid shape correspondence
using random forests. In CVPR. IEEE, 2014.
[22] J. Shotton, A. Fitzgibbon, M. Cook, and A. Blake.
Real-time Human Pose Recognition in Parts from Single Depth Images . In CVPR. IEEE, 2011.
[23] J. Starck and A. Hilton. Correspondence labelling for
wide-timeframe free-form surface matching. In ICCV.
IEEE, 2007.
[24] M. Straka, S. Hauswiesner, M. Ruether, and
H. Bischof. Skeletal graph based human pose estimation in real-time. In BMVC, 2011.
[25] M. Straka, S. Hauswiesner, M. R¨uther, and H. Bischof.
Simultaneous shape and pose adaption of articulated
models using linear optimization. In ECCV, 2012.
[26] R. W. Sumner and J. Popovi´c. Deformation transfer
for triangle meshes. In ACM Transactions on Graphics (TOG). ACM, 2004.
[27] J. Taylor, J. Shotton, T. Sharp, and A. Fitzgibbon. The
vitruvian manifold: Inferring dense correspondences
for one-shot human pose estimation. In CVPR, 2012.
[28] F. Tombari, S. Salti, and L. Di Stefano. Unique signatures of histograms for local surface description. In
ECCV. Springer, 2010.
[29] D. Vlasic, I. Baran, W. Matusik, and J. Popovi´c. Articulated mesh animation from multi-view silhouettes.
In TOG. ACM, 2008.
[30] A. Zaharescu, E. Boyer, and R. Horaud. Keypoints and
local descriptors of scalar functions on 2d manifolds.
IJCV, 2012.